Variables Selection for Multiclass SVM Using the Multiclass Radius Margin Bound

Support vector machines (SVM) are considered as a powerful tool for classification which demonstrate great performances in various fields. Presented for the first time for binary problems, SVMs have been extended in several ways to multiclass case with good results in practice. However, the existence of noise or redundant variables can reduce their performances, where the need for a selection of variables. In this work, we are interested in determining the relevant explanatory variables for an SVM model in the case of multiclass discrimination (MSVM). The criterion proposed here consist in determining such variables using one of the upper bounds of generalization error specific to MSVM models known as radius margin bound [1]. A score derived from this bound will establish the order of relevance of variables, then, the selection of optimal subset will be done using forward method. The experiments are conducted on simulated and real data, and some results are compared with those of other methods of variable selection by MSVM.